Improvement of Prostate Cancer Aggressiveness Prediction Performance Using a Self-Supervised Learning Model Fine-Turned on Similar Medical Images from Multi-Parametric MR Images

Abstract

In this paper, we propose a prostate cancer aggressiveness prediction model using self-supervised learning based on SimCLR with multi-parametric MR images. Self-supervised learning model is initially trained on the STL10 dataset, and then fine-tuned on the ProstateX dataset, which is similar to the downstream task dataset. To predict prostate cancer aggressiveness, downstream tasks are performed using each sequence of images from the multi-parametric MR dataset. The predicted results are combined using either majority voting or average voting for ensembling. Experimental results demonstrate that the self-supervised learning model fine-turned with similar images improves the performance by an average of 4.56% in accuracy, 20.69% in sensitivity, and 12.02% in negative predictive value. The ensemble method using majority voting with the self-supervised learning model fine-turned on similar images from the multi-parametric MR dataset yields the best performance in terms of accuracy (72.58%), balance accuracy (72.16%), and sensitivity (67.86%).